Neural network machine learning model

Abstract

Certain aspects provide a method for assigning a plurality of physical properties in space and time of a target underground region to a plurality of structural nodes defined for a first layer of a graph neural network machine learning model; constructing a regular grid from the plurality of structural nodes; generating, by a neural operator layer of the graph neural network machine learning model and using a fast Fourier transform, a neural operator output; projecting, by the neural operator layer via an inverse fast Fourier transform, the neural operator output onto the regular grid to generate an inverse grid; and generating a prediction from a second layer of the graph neural network machine learning model.

What is claimed is: 
     
         1 . A method comprising:
 assigning a plurality of physical properties in space and time of a target underground region to a plurality of structural nodes defined for a first layer of a graph neural network machine learning model;   constructing a regular grid from the plurality of structural nodes;   generating, by a neural operator layer of the graph neural network machine learning model and using a fast Fourier transform, a neural operator output;   projecting, by the neural operator layer via an inverse fast Fourier transform, the neural operator output onto the regular grid to generate an inverse grid; and   generating a prediction from a second layer of the graph neural network machine learning model.   
     
     
         2 . The method of  claim 1 , wherein the regular grid is generated as an output of the first layer of the graph neural network machine learning model. 
     
     
         3 . The method of  claim 2 , wherein the first layer takes, as input, the plurality of structural nodes. 
     
     
         4 . The method of  claim 3 , wherein the regular grid expands the plurality of physical properties into a latent space. 
     
     
         5 . The method of  claim 1 , wherein the neural operator layer takes, as input, the latent space of the plurality of physical properties. 
     
     
         6 . The method of  claim 1 , wherein the second layer takes the inverse grid as input. 
     
     
         7 . The method of  claim 1 , wherein the plurality of physical properties is from a 3D reservoir model with arbitrary geometry and includes porosity, permeability to flow in multiple directions, rock properties, depth, and well locations. 
     
     
         8 . A system comprising:
 a memory comprising computer-executable instructions; and   a processor configured to execute the computer-executable instructions and cause the system to:
 assign a plurality of physical properties in space and time of a target underground region to a plurality of structural nodes defined for a first layer of a graph neural network machine learning model; 
 construct a regular grid from the plurality of structural nodes; 
 generate, by a neural operator layer of the graph neural network machine learning model and using a fast Fourier transform, a neural operator output; 
 project, by the neural operator layer via an inverse fast Fourier transform, the neural operator output onto the regular grid to generate an inverse grid; and 
 generate a prediction from a second layer of the graph neural network machine learning model. 
   
     
     
         9 . The system of  claim 8 , wherein the regular grid is generated as an output of the first layer of the graph neural network machine learning model. 
     
     
         10 . The system of  claim 9 , wherein the first layer takes, as input, the plurality of structural nodes. 
     
     
         11 . The system of  claim 10 , wherein the regular grid expands the plurality of physical properties into a latent space. 
     
     
         12 . The system of  claim 8 , wherein the neural operator layer takes, as input, the latent space of the plurality of physical properties. 
     
     
         13 . The system of  claim 8 , wherein the second layer takes the inverse grid as input. 
     
     
         14 . The system of  claim 8 , wherein the plurality of physical properties is from a 3D reservoir model with arbitrary geometry and includes porosity, permeability to flow in multiple directions, rock properties, depth, and well locations. 
     
     
         15 . A non-transitory computer-readable medium comprising instructions that, when executed by one or more processors of a computer system, cause the computing system to perform operations, the operations comprising:
 assigning a plurality of physical properties in space and time of a target underground region to a plurality of structural nodes defined for a first layer of a graph neural network machine learning model;   constructing a regular grid from the plurality of structural nodes;   generating, by a neural operator layer of the graph neural network machine learning model and using a fast Fourier transform, a neural operator output;   projecting, by the neural operator layer via an inverse fast Fourier transform, the neural operator output onto the regular grid to generate an inverse grid; and   generating a prediction from a second layer of the graph neural network machine learning model.   
     
     
         16 . The non-transitory computer-readable medium of  claim 15 , wherein the regular grid is generated as an output of the first layer of the graph neural network machine learning model. 
     
     
         17 . The non-transitory computer-readable medium of  claim 16 , wherein the first layer takes, as input, the plurality of structural nodes. 
     
     
         18 . The non-transitory computer-readable medium of  claim 17 , wherein the regular grid expands the plurality of physical properties into a latent space. 
     
     
         19 . The non-transitory computer-readable medium of  claim 15 , wherein the neural operator layer takes, as input, the latent space of the plurality of physical properties. 
     
     
         20 . The non-transitory computer-readable medium of  claim 15 , wherein the second layer takes the inverse grid as input.